A presentation providing a case for the applicability of recent developments in AI, applied in medicine, to asset management. The particular example discussed is the prediction of machine failure.
Call Girls In Attibele ☎ 7737669865 🥵 Book Your One night Stand
Progress in AI and its application to Asset Management.pptx
1. Progress in AI and its Application
to Asset Management
Derryn Knife
2. Survival Analysis
• Survival analysis is the broad set of
tools and techniques to determine the
“time to an event”.
• Essentially answers questions about
how long something will last.
• Asset managers can use survival
analysis to answer questions about
stocking strategies, maintenance
strategies since we can understand
how long a piece of equipment will
function.
3. Application of Survival Analysis to AM
Typical Question: “When will my item fail?”
Solution: Use Non-Parametric estimates.
Complication: Failures only so high…
Non-parametric cannot be used extrapolate beyond
highest failure.
Solution: Use Parametric estimation
Complication: What about the
thickness/temperature/other single factor that
affects this?
Simple distributions cannot account for single factor
impacts.
4. Application of Survival Analysis to AM
Solution: User Accelerated Life / Accelerated Failure Time /
Proportional Hazards Model.
Complication: But some of these vary over time?
Simple PH/ALT/AFT models cannot account for time-varying
coefficients.
Solution: Use Time-Varying Proportional Hazards (or ALT, AFT).
Complication: But how do I know when the time varying
covariates change…
Time-Varying models cannot predict when the varying
coefficients will change. Great to answer questions like “what is
at most risk?” But not when.
Solution:…
Machine Learning used in Medicine
5. What is machine learning?
Supervised
• Takes a known answer.
• Algorithm attempts to ‘regress’
onto the known answer with the
data given.
• Linear regression is a simple
version of supervised learning.
• Applied anywhere where a
predicted value is needed.
Unsupervised
• No known output, just the data.
• Used to learn representations of
the data.
• Representations can be used to
reduce the size of data.
• Also used to reduce varied sizes
to a common size.
• Applied for outlier detection,
recommendations, clustering.
6. What is Deep Learning?
• Deep learning is a subset of machine
learning. Much of the excitement
about AI focuses on deep learning.
• Conceptually it aims to replicate the
structure of a neuron in the human
brain. It takes multiple inputs, and
‘activates’ at particular a threshold.
• Notable since it can continue to
consume larger data sets where other
ML algorithms reach a plateau.
• Many different and flexible
‘architectures’. Deep Learning vs Other ML (Alom et al.)
7. Conventional Approach to RUL – Regression
With a regression approach the Machine Learning Algorithm attempts to
learn one number. Usually from just time series data, not disparate types.
The problem with this approach is that not many decisions can practically
be made. It cannot answer “what is probability of failing?”
100
hours
8. Next usual approach – Logistic Regresssion
Alternately, one can regress on whether the event will or will not happen
within a given window of time.
Although this does give you the risk it is not good for long term forecasts
since it cannot predict where within a longer window.
10%
(chance within
100 hours)
9. Recent Progress in AI
Recent progress in AI has seen Deep Learning algorithms output not just a single
value, but a distribution. By doing this, these new approaches overcome many of
the short falls of traditional ML and combines it with the power of survival
analysis.
Predicting a failure distribution allows one to understand risk over time. This
allows for optimal decision making for intervention.
10. Recent Progress in AI – An Example
• Has been applied in the medical
literature.
• A particularly notable example is
Shu et al. They use feature
extraction on x-rays and vital
sign time series data.
• They then concatenate these
reductions with demographic
data. This vector is then fed into
a neural network to predict risk
(a distribution).
11. Applicability to Asset Management
[0.36401788, 0.72096758, … 0.65392181, 0.22536759]
[0.05062463, 0.01495256, …
0.78682389, 0.24272541]
[0.05062463, 0.01495256, …
0.78682389, 0.24272541]
[0.05062463, 0.01495256, …
0.78682389, 0.24272541]
Clinical notes are similar to investigations or the notes of a
maintenance history. Can reduce this to a numeric
representation.
Input characteristics, ore
type, water characteristics,
electrical power profile etc.
can be reduce to a numeric
representation
Duty profile, throttle, on/off
commands, number of
cycles etc can be reduced to
a numeric representation.
[0.05062463, 0.01495256, …
0.78682389, 0.24272541]
Machine state,
characteristics even a time
series of photos can be
reduced to a numeric
representation.
12. Applicability to Asset Management
Environment
Duty / Task
Inputs
Maintenance
Equipment
This approach can answer questions about when the equipment will
fail based on many combined factors, including time varying factors.
What if these change?
Answer: Simply change the vector! The result is an extremely
powerful method of prediction for optimisation and changing
circumstances
Notes de l'éditeur
The purpose of this presentation is to argue, and hopefully convince, the Asset Management community that there are cutting edge solutions, designed to solve issues in medicine, that can be immediately applied to issues experienced by asset owners.